x-transformers 1.37.5__tar.gz → 1.37.6__tar.gz

{x_transformers-1.37.5/x_transformers.egg-info → x_transformers-1.37.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.37.5
+Version: 1.37.6
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.37.5 → x_transformers-1.37.6}/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.37.5',
+  version = '1.37.6',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

{x_transformers-1.37.5 → x_transformers-1.37.6}/x_transformers/attend.py

@@ -64,15 +64,15 @@ print_once = once(print)
 
 # alternative distance functions
 
-def qk_l2_distance(q, k):
+def qk_l2_dist_squared(q, k):
     if k.ndim == 3:
         k = repeat(k, 'b j d -> b h j d', h = q.shape[1])
 
     q, packed_shape = pack_one(q, '* i d')
     k, _ = pack_one(k, '* j d')
 
-    distance = torch.cdist(q, k)
-    return unpack_one(distance, packed_shape, '* i j')
+    l2_dist_squared = torch.cdist(q, k) ** 2
+    return unpack_one(l2_dist_squared, packed_shape, '* i j')
 
 # functions for creating causal mask
 # need a special one for onnx cpu (no support for .triu)
@@ -353,7 +353,7 @@ class Attend(Module):
         if not self.l2_distance:
             sim = einsum(f'b h i d, {kv_einsum_eq} -> b h i j', q, k)
         else:
-            sim = -qk_l2_distance(q, k)
+            sim = -qk_l2_dist_squared(q, k)
 
         sim = sim * scale
 

{x_transformers-1.37.5 → x_transformers-1.37.6/x_transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.37.5
+Version: 1.37.6
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang